(navigation image)
Home American Libraries | Canadian Libraries | Universal Library | Community Texts | Project Gutenberg | Children's Library | Biodiversity Heritage Library | Additional Collections
Search: Advanced Search
Anonymous User (login or join us)
Upload
See other formats

Full text of "Leonard Kleinrock Interview. 3"

An Interview with 
LEONARD KLEINROCK 
OH 190 
Conducted by Judy O'Neill 
on 
3 April 1990 
Los Angeles, CA 
Charles Babbage Institute 
The Center for the History of Information Processing 
University of Minnesota, Minneapolis 
Copyright, Charles Babbage Institute 
------------------------------<page break>-----------------------------
nicely. We were the f'zrst node, and then in the next three months three more nodes were added - we had a four- 
node network by December, the famous four-node network. Here is the book I was talking about earlier 
[Communication Nets]; it eventually went into paperback. In this other book of mine [Queuing Systems, Volume 
II: ComœuterAœœlications], we see the network that was initially implemented, namely, that four-node network 
[p. 306]. And then it began to grow;, by the middle of 1970 it looked like this. Early in 1971 it began to grow 
again. And all the while we were making measurements. And they were not just passive measurements - we 
did more than observe what was going; we were experimenting and generating artificial traffic. We became the 
network measurement center. Meanwhile my research quickly moved into networks, which I was waiting to do. 
I began publishing a great deal on the subject. 
O'NEILL: Okay. I have some questions on some of the things that you have mentioned. I have a pretty good 
overview up to the point of your early ARPANET work. You talked about your advisor, Edward Arthur, who 
was working on a classified project that you didn't have access to for one of the services. Do you know if he 
took your work and applied it to the network? 
KLEINROCK: I don't know, and I sort of doubt it, because he wasn't very heavily coupled into that. I think 
he attended a conference in which these problems came up, and I don't recall him being heavily coupled. I do 
remember that there was a thing called UNICOM, I believe, that became interesting in the early to mid 1960s. 
And I was at a conference in the early 1960s where that was mentioned, and my ideas were made known there. 
I never worked on that project. 
O'NEILL: When you decided to go to UCLA, did you have any interest in or knowledge of the DARPA 
projects that the Western Data Processing Center had gotten started? 
KLEINROCK: No, but that's where Ivan enters the picture. I believe it was 1965, or so, that he came by with 
his ARPA hat and suggested that we network the three IBM 7090s on campus. There was one in the campus 
------------------------------<page break>-----------------------------
computing facility, one at Western Data Processing Center, and one over in the medical center. In 1967, Larry 
entered the picture and began talking about his plans to fund a network implementation. In April 1969 I 
received an ARPA research contract. It was obvious I was going to take a key role in the forthcoming 
ARPANET. So I began as principal investigator in April 1969. We still have a major contract with DARPA. 
O'NEILL: There is a Program Plan in the DARPA files for the Western Data Processing Center, but it is dated 
1963. 
KLEINROCK: That was a separate thing. They were providing computing services to some set of entities. 
O'NEILL: When you were at Lincolrb or I guess it would have been after you had already come out here, Larry 
Roberts and Tom Marill did an experiment connecting the TX2 and the Q32. Were you aware of that work? 
KLEINROCK: Oh, very much so. Larry spoke to me quite a bit about the problems they were having and the 
way he was trying to solve it. I was aware of but not participating in that experiment. I know that is what 
triggered Larry to try to get a system that would work more smoothly. The interesting thing is, as I recall, that 
part of the motivation for tl network is the fact that in 1967, in the mid 1960s DARPA was heavily supporting 
a lot of people doing work on time-sharing. And every time an investigator got a new contract, the first thing 
he wanted was a computer - the best and biggest. Pretty soon Larry said, 'Itds is getting ridiculous,' because 
each facility they created evolved into a specialized kind of facility, like the graphics capability at Utah, the 
database capability at SRI, and the simulation capability at UCLA. So Larry came up with the concept of a 
resource sharing network, where there would be specialized sites, and if you wanted that special capability, you 
connect to that site to get it, or you would pull back data or programs and use them locally. That was one of 
his motivating reasons, namely, to reduce the number of time-sharing systems he had to support. 
O'NEILL: Were you also keeping track of what Paul Baran was doing in his Rand reports? 
------------------------------<page break>-----------------------------
KLEINROCK: I was well aware of his results. In fact I quoted his results in my own dissertation. He had 
some of the important ideas of packet switchin E as a system. The thin E that really drove my own research was 
the idea of a message switching network, which was a precursor to the packet switching networks. The 
mathematical tool that had been developed in queueing theory, namely queueLug networks, matched perfectly 
the model of computer networks. Actually, it didn't match perfectly and I had to adjust that model to fit the 
realities of computer networks. Then I developed some design procedures as well for optimal capacity 
assignment, routing procedures and topology design. 
O'NEILL: I'm trying to pull it all together. It is clear Larry Roberts had a big part in tl, but it is important 
to understand historically where the roots are and how it came together. 
KLEINROCK: Well, you know, Donald Davies of the National Physical Laboratory in the UK was also a key 
contributor. Donald Davies had the same idea of a packet network independent of Larry Roberts. Both of them 
post-dated Paul's basic ideas. Now, Paul didn't have the total picture in mind, but he had the essentials there. 
Unfortunately, Paul's work was not properly recogni:ed in the early literature. His work was in 1964, I believe. 
O'NEILL: The reports were actually published in 1964, but they had been written before that and they had had 
some exposure before that. 
KLEINROCK: And I was aware of his work, but again, it didn't play a major part in my thinklag until some 
years later. But I would credit him with the first ideas. Larry and Donald Davies worked in parallel and 
independently. Unfortunately Donald Davies at NPL never got the proper funding to really let that system 
blossom physically. I believe he made a one-node switch at NPL. He had the ideas though. He was definitely 
a forward thinker. Larry fortunately had the resources to make networking happen. 
6 
------------------------------<page break>-----------------------------
O'NEILL: You said that Larry brought you and others out to Washington to start talking about the specifications 
for the network. Was that after the principal investigators meeting where he had talked to the host community? 
I'm assuming you weren't a principal investigator because you said you hadn't gotten involved until later. 
KLEINROCK: I attended,.but later than 1967. I attended one in late 1968 or early 1969. The meeting was in, 
of all places, Alta, Utah - a wonderful ski resort. 
O'NEILL: So at some point in 1967, you don't recall just when, you went to a meeting in Washington. 
KLEINROCK: I remember some of the people there. I remember that Tom Cheatham was there. I believe 
Herb Baskin was there also, and I think it was he, or it may have been Doug Engelbart, who was hot on time- 
sharing and said, 'If thi. network can't give me a half a second response time, then I can't use it for time- 
sharing.' So we specified that there should be a half a second response time. We soon found out it could give 
two tenths of a second response time using the 56 KBPS lines we had. But it was there that it was decided. And 
I banged my fist on the table and said, %Ve've got to put measurement software in there. If this is going to be 
an experiment, we must make measurements. So it evolved; we decided on a reliability criterion, and specified 
the format and the nature of thi. network. 
O'NEILL: What was the total size of this group that was together in Washington? 
KLEINROCK: Ten... on that order. 
O'NEILI,: Okay. Was this an ongoing series of meetings, or was this just a couple of days? 
KLEINROCK: One or two days, I can't remember which. Just one shot. 
------------------------------<page break>-----------------------------
O'NEILL: Had there been a document prepared for that meeting? Were you working from a draft proposal? 
KLFINROCK: No. What happened thereafter I had no access to, but a formal RFQ spec was prepared. 
O'NEILL: Okay, so was that your only direct involvement prior to actually getting the IMP installed? 
KLEINROCK: No, I believe I was asked to help evaluate some of the proposals as well, but in an unofficial 
way. 
O'NEILL: So you weren't part of a committee evaluating the proposals? 
KLEINROCK: No. I was aware of the numbers that were coming in. I was aware that many of them suggested 
the same computer for the switch, namely, the Honeywell DDP-516. 
O'NEII.I.: In the proposals that you happened to see, was there a lot of variation in how they proposed to do 
it? 
KLEINROCK: I only remember two of them in some detail. I don't remember how much software they 
specified. And I don't remember, frankly, when the concept of distributed control routing was described in 
detail. We suggested it, as opposed to centralized control. 
O'NEILL: We, being the group in Washington? 
KLEINROCK: Yes. It was clear that we wanted no central location which if it failed, would take down the 
network. We distributed control throughout. But that is the concept; now the particular algorithm I believe was 
developed at BBN. You may or may not know that BBN was extremely protective of the software. They wrote 
------------------------------<page break>-----------------------------
the software accordinv to the gross spec that we gave. And they also gave us the network algorithms, but they 
wouldn't let us look at the code. Now there we were at UCLA; it was our job to test and try to break the 
network - to experiment with it. And when we looked at the algorithm, we told BBN that there were some 
problems. There would be loops in the routing procedure and some deadlocks in the flow control algorithm. 
And they were very reluctant to make changes. They saw us really as a nuisance, for good reason. They had 
a job to do and we were telling them to do it differently. But the interesting thing is they wouldn't let us see 
the source code. Even though we suspected a problem in their implementation, we could never see it. We told 
them, Iere is a problem, fix it.' They said it would take six months to f'tx, so they never did. Finally there was 
pressure put on them to make the software public (I forget where that pressure came from - it was either 
DARPA or an external community that said, q_ok, the government paid for the software; you don't own it.') 
Finally they were forced to open up the software. Then we had access to the code, and what we did then was 
to find the trouble and not only did we tell them to fix it, but we also told them how to fix it. It still took six 
months. (Laugh) 
O'NEILL: When the first IMP was installed here at UCLA, were you personally involved in getting it up and 
Iqnning? 
KLEINROCK: Oh, very much so. It cost me a great deal in fact. I was involved with a start-up company at 
the time, and my participation was severely diminished as a restfit of that. 
O'NEILL: Was that the Linkabit Corporation? 
KLEINROCK: Yes. I have since formed another company on my own. 
O'NEILL: The Technology Transfer Institute. In fact, I was curious about that name. It kind of invites a 
question as to what technology, and who is it transferred to? 
------------------------------<page break>-----------------------------
KLEINROCK: Okay. The concept of that name is that we were going to transfer technology via seminars, that 
is, to get the word out there quickly. If you want to get to that now, we can. Or maybe you want to wait. 
O'NEILL: Why don't we just go ahead, as long as we're talking about it. 
KLEINROCK: I published a two-volume book in the mid 1970s. The first was on queueing theory. The second 
one was on computer applications. It was the second book that talked about computer networks. There were 
two major chapters on it. It included much of the material here [in the 1964 monograph] and a lot of the work 
I had done since. This came out in 1976, so that's seven years after the ARPANET began. There was a lot 
about the ARPANET technology and its development in the intervening years. The book was new, the topic 
was hot, and my name was known. I thought, v'hy not do a seminar to get this information out quickly.' 
Everybody was pushing me to do that. So in the summer of 1976, my wife and I formed a little company, created 
a 'Computer Network' seminar, publicized the seminar, and ran the same seminar in Dallas, Washington, and 
L.A. It was a big success. So I formed this company and said, 'Let's go.' And what was the reason for doing 
this? It's that I had access to the best talent in the performance evaluation and networking area in the country. 
People like Larry Roberts, like Howie Frank, like Norm Abramson, like Peter Denning, and myself. Using them 
as lecturers, I created five seminars, five separate seminars, and we launched that in the winter of 1977. We 
repeated that twice a year; then it went to six seminars twice a year. And it was going along f'me, very nicely. 
Then just to continue the history now, in the spring of 1979, we went from six twice a year to twelve twice a year. 
And more significantly, we took on James Martin. Do you know James Martin's name? 
. O'NEILL: Yes. 
KLEINROCK: Okay. He had been at IBM. In 1977 he took a sabbatical, got some friends of his in the U.K:, 
and created a five-day seminar; they administered and he spoke - around the world (fourteen places in the 
world). It was a big success. In 1978 they did it again, and these people in the U.K. were getting tired; so they 
10 
------------------------------<page break>-----------------------------
decided to parcel out the world. My company was thriving and Jim and I had known each other very well since 
1970; we had done some seminars in Israel and Europe together. TrI took over the Jmes Martin seminar in 
North America. And it has beea so ever since. The number of seminars has grown. I was the president of the 
company at the time; I since got someone to take over that role. I don't handle the day-to-day operations. The 
thin which made it work are as follows. First of all I had the network of the best talent. I knew who the best 
people were. And I could evaluate them. They had to satisfy two of three criteria, every speaker: they either 
have written the key text in the field, and be the pioneer in the field, and be an excellent speaker - any two of 
those three criteria. The second key was the synergistic relationship between my teaching, my research, and the 
thin that I was teaching out in the seminars. It was all the same kind of thing, so it was low overhead to me 
to create these seminars. It was what I was doing anyway. That's one of the reasons it worked. It is the reason 
I didn't have to leave the urdversity to make it happen. I got the right management in, and I controlled the 
quality. And it is still going great. 
O'NEILL: When you first started this, was there any concern by, or connection at all, with DARPA, the fact 
that they were providing money to your networking research? 
KLEINROCK. No. In fact, they were very much interested in and happy to have spinoff companies, and to get 
that information out there. 
O'NEILL: Was there anything you would consider active encouragement of that? 
KLEINROCK: No. But a variety of the people at DARPA taught for me. Like Larry, like Bob Kahn like Vint 
Cerf. They were people at the office at various times. And many of the people they were supporting were 
teaching for me - Dave Farbet, Howie Frank, Carl Sunshine, John McQuillan, etc. But there was no active 
support, no. 
11 
------------------------------<page break>-----------------------------
TAPE 1/SIDE 2 
O'NEILL: Let's talk a bit more about the situation here at UCLA. I had gotten the impression that there were 
a lot of students coming in working on the network, doing research on the network. Was that something that 
you actively set up, tried to work it that way? 
KLEINROCK: Very much so. And it had to do with the fact that I had this large ARPA contract that was 
supporting the things I was interested in, but also helped point the direction as to where the research should go. 
For example, there were phases of research that I passed through. I started doing some queuehag theory, some 
time-shared modeling, and then moved into networking with support from DARPA. The ARPANET was the 
first piece of that. There were various waves of research that occurred from there on. Wave one was the 
ARPANET, for me. Now, of course, the time-sharing work was already being supported by DARPA elsewhere. 
I was not being supported at the time, but I was heavily involved in that research. All of the money I received 
was used to support Ph.D. students. But also a big piece of the money that I was given by DARPA was used 
to support the implementation. You see, UCLA was in charge of creating the host-to-host software. Steve 
Crocker, Vint Cerf and Jon Postel all worked for me at the time. I had at one time about forty people - 
secretaries, programmers, managers, faculty and students. It was amazing, I started this huge group. And Steve 
Crocker was in charge of the software side of it. I had other people in charge of the physical side of the facility, 
our host, our IMP, and all the rest. But Steve Crocker played a very important role on the software side. Steve 
was getting his Ph.D.; so was Vint and so was Jon. I f'filed my students full of networking ideas. Now, as I say, 
we became the ARPANET measurement center as well. Around that same time DARPA decided to create a 
transadantic link to Europe via satellite. 
O'NEILL: This was 19727 
12 
------------------------------<page break>-----------------------------
KLEINROCK: 1972, 1973 in that range, yes. And we decided to experiment with what is called a broadcast 
satellite channel using ALOHA as the access scheme. So we began to study packet satellite switching. And 
some of my students started to move into that research area. For example, one of my students was Mario Gerla 
who later worked for Howie Frank. We hired him back from Network Analysis Corporation. I'd say Mario's 
name was the most sioificant one that came out of the ARPANET research. Then came the satellite stuff. 
Simon Lain is a name you may or may not know. He is now at the University of Texas. His thesis was studying 
ALOHA as an access scheme mathematically. And that came out of the satellite switching meetings that we had. 
There was a set of meetings that went on to develop this technology. I brought Simon with me to these meetlng 
to act as my right-hand man. Out of that came his dissertation. Then ARPA started moving to packet radio 
on a metropolitan area networking basis. The application was to a soldier rmning arotmd the field with digital 
radios, or possibly tank moving across a batfiefield; that led to the SURAN, survival radio network, and the 
whole packet radio project began. We started studying CSMA, carrier sense multiple access, which eventually 
led to ETHERNET. Fouad Tobagi was also my student; he's now at Stanford. For his thesis, I told him 
'You're going to evaluate CSMA.' He did a beautiful job. There is a sequence there: Gerla, Lain, and Tobagi. 
All at universities, all well known, all right now. Following that came some work on local area networks that 
DARPA didn't really promote heavily. But it was a technology that came out of the packet radio project. The 
same ideas for sharing a ground radio network are the same ideas you need to share a bus or-a ring-type 
communications medium. Bob Metcalf, who was present in all of this (he was up at Xerox at the time) came 
up with the idea of CSMA/CD, with collision detect to use on a wire, and out of that, of course, came Ethernet 
and all the rest followed. Some other of my students' names may interest you. Bill Naylor who was one of the 
programmers under Steve Crocker, did a dissertation for me. His job was to do the measurement of the 
ARIANET. He and a fellow named Gerry Cole were doing measurements theses. They took data and they 
created and analyzed models. Out of the packet radio came a young man named John Silvester. He is well 
known in the packet radio research field. He is over at USC right now. And there are many others. Do you 
care about these names? 
------------------------------<page break>-----------------------------
we did the ARPANET measurement. Following that we were tasked with doing some satellite measurement 
and some packet radio measurement. I spun off the packet radio measurement as well. So there were two 
spawns that took place out of the original ARPA support. 
O'NEILL: Were you free to make these subcontracts? 
KLEINROCK: No, no, no. I did it formally through the ARPA office. I said, q_ok you should be supporting 
these guys on their own.  
O'NEILL: And did that take away from the money you were getting? 
KLEINROCK: Well, my pile was growing. It was growing bigger than I wanted. It was approaching a million 
dollars a year. I don't need that much money. As it is now, I'm being supported at about two thirds of a 
million, and that is probably going to stop come next renewal, because they can't afford so much money for 
'theory and performance evaluation,' which is unfortunate. I'm going to shrink considerably this next round. 
Right now I'm supporting seven Ph.D. students and a couple of secretaries. That will go way down. A lot of 
it has to do with new structure in the ARPA office, which has totally changed. 
O'NEILL: That brings up another point, the changes in the ARPA office over time. Obviously you have seen 
that change over several years now. Maybe we should back up, and why don't you describe what your interaction 
was like with the office initially. 
KLEINROCK: Okay. The reason I got the contract initially was because Larry and I knew each other well, we 
respected each other, and the ARPANET thing got going. At that time he was not the head of the office, Bob 
Taylor was. Then Bob left and shortly thereafter Larry took over. Things with DARPA were beautiful at that 
time. After Larry left, Bob Kahn eventually moved in as the head of IPTO. 
------------------------------<page break>-----------------------------
O'NEILL: Licidider came back for a short time, and then there was Russell, and then Kahn. 
KLEINROCK: True. Kahn had a very long tenure, and Bob and I got along fatuously. In fact, Bob was in the 
initial ARPANET experiment, as you well know. He came out here as a BBN rep, and did some of the initial 
measurements with me. We had excellent rapport. Eventually, Bob Kahn and Bob Cooper had some problems 
regarding support of networking. 
O'NEILL: Who is Bob Cooper? 
KLEINROCK: Bob Cooper was the head of all of DARPA toward the end of Bob Kahn's tenure. 
O'NEILL: Where do you look now for funding? 
KLEINROCK: Me personally? I'll probably not look for large contracts. I think I'll be happy with a smaller 
amount and fewer students. It is a big administrative drain to run a large contract. I've been running six, seven, 
eight students continually, and I may just shrink down to three or four, reduce my staff and just manage with 
the money I have. Somehow, it is probably time to cut it down. 
O'NEILL: But that would still be ARPA money then that you would run on? 
KLEINROCK: Yes, to the ARPA money. I might also go to NSF or I might go to IBM, if I feel I need some 
more, but I probably won't. 
O'NEILL: If you had wanted to maintain your level, are NSF and IBM the places you would go? 
KLEINROCK: Possibly. 
16 
------------------------------<page break>-----------------------------
O'NEILL: I would like to back up a little bit, again, to the working relationship when the ARPANET was being 
installed. You mentioned your inter-relationship with BBN. Can you talk more about that? Not only BBN but 
obviously the Network Analysis Corporation, and SRI perhaps. 
KLEINROCK: Okay. Let me talk about NAC first. That goes back to this picture again (in Volume II). In 
the early days there was no problem in designing the topology of the ARPA network. I mean, how else would 
you connect four nodes? And maybe the same for ten nodes where this group is on the east coast and this group 
is on the west coast - it is pretty obvious. As we got into larger networks, topology became a major issue. I had 
known Howie Frank for many years prior to the whole ARPA thing. In fact I met him in 1966 when I did a 
lecture for him at Berkeley - he was on the faculty at Berkeley at the time - and so was Ivan Frisch. They had 
written this very important book on network flow. In fact, I almost got killed going to lecture for them up in 
Berkeley at the time. I was coming out of Sequoia in a station wagon and the brakes failed going downhill at 
forty-five miles an hour... but, that's another story. (Laugh) Actually, in 1968, Howard Frank and myself, and 
a bunch of other guys, including Ivan Frisck, went to Washington, D.C. to work for the Executive Office of the 
President. 
O'NEILL: So you were part of the same group that Howard Frank was involved with? 
KLEINROCK: Half and half. Do you know about that whole effort? A guy named David Rosenbaum was the 
one who promoted this thing, and he knew Howie. Basically, he got the government to agree to create this 
group, that summer, to try to show how network flow theory could be useful to government problems. I was 
not on the full-time team. I spent two weeks there. The problem they chose to work on was a problem for the 
Federal Power Commission on desig'ning natural gas pipeline networks. Great work! Very important algorithms 
came out! And the most important person there, to be honest with you, is none of the names I mentioned - was 
a guy named Danny Kleitman, who is a mathematics professor now at MIT. Brilliant guy and he came up with 
the key algorithm. We made some important progress. We saved the government $20,000,000 dollars over a 
17 
------------------------------<page break>-----------------------------
few years, with a small investment. Shortly thereafter Howie, Ivan and Dave formed the Network Analysis 
Corporation. 
I went there to lecture a couple of times. But I knew the whole group. In about 1971 or so Larry was at my 
house, and I suggested that he meet Howie Frank to assist in the topology design problem. So I put them 
together. And it was a click. Then Larry gave NAC the contract for doing the topological design for the ARPA 
network. That's the history on that, and it's continued ever since. UCLA was the first node, and SRI was the 
second node. So this was the first link in the ARPA network - there it is, there it is, and there it isn't. Around 
this time we were busy simulating the network and also measuring it, and we found that there was almost no 
reason to have that link in there. Independent of that Howie Frank's group came in, and they did a design, and 
they also felt that that link was unnecessary. So you see, that takes some daring, bemuse if you mess around, 
if you make a mistake you are in trouble. At this point some intelligence was built into the network. Howie 
always made the point that network flow theory is a really complicated theory and you can't do network flow 
design on the back of an envelope. If you look at this network, this network looks like the back of an envelope. 
(Laugh) So NAC and we at UCLA were very friendly, professionally, personally;, but we also were competitive, 
because one of my students, essentially Mario Gerla and myself, were busy doing network design, and developing 
network design procedures. I had done some, and Mario and I developed some more. And we developed a 
network design algorithm called CBE, the Concave Branch Elimination method. NAC had theirs, called the Cut 
Saturation Method. They were competitive, and they were both effective. But theirs was the one that was used 
because they were under contract. We later showed, through some research, that even though they came up with 
a very different network, it had about the same cost performance profile as networks we were generating, which 
is rather interesting. So we were competitive but very cooperative. Mario then went to work for NAC for about 
three or four years. He worked with Wushow Chou, who is now at North Carolina State University. We then 
hired Mario back here, after he had seasoned out there. So that was one of our relationships. 
18 
------------------------------<page break>-----------------------------
Now for BBN. I characterize the BBN relationship this way. BBN, as were most software groups at that time, 
were largely uainterested in performance. That's an exaggeration. But by and large a programmer, a software 
guy, simply wants to get a piece of software that works. That's hard enough. Whether it works efficiently or 
well is not usually the issue; they worry about that maybe way down the llne. So BBN was not focusing on 
efficiency or performance but on making it work and worrying about the logic. So we had a rmnlng battle with 
them in terms of showing them how to improve performance. Another student of mine (Gary Fultt) and I 
showed them early on that their routing procedure was inferior to two others we developed. We did this in 1970. 
Finally in 1979 they implemented something very much like one of our procedures and then some years later 
another piece of it. BBN monitored the network. They measured traf/ic, and they measured llne failures, and 
/MP failures. We were doing things like measuring response times (which they were not doing) and stressing 
the network by generating extra network trafiic to see what would happen. We would, almost on demand, break 
the network. There were faults in the software, logic faults. So you can see BBN was not very happy with us 
showing up their faults and telhag them to fix them. Sooner or later they did. Any large system is going to 
contaia faults. But we became experts in f'mdin these things, and I document a whole bunch of them in here. 
One time we were makin E measurements, which means we turned on the IMP software which made 
measurements and which were sent back to us at the Network Measurement Center. As a result of that the 
network kept crashing. So BBN called us up and said, 'You're crashing the network!' And we said, %Ve need 
to make measurements.' And we found out why. Are you a technical person at all? 
O'NEILL: Yes, somewhat. 
KLEINROCK; Well, the IMP had only so much storage, initially a very limited amount of storage. And some 
of it was set aside to reassemble packets into messages - the reassembly function. Specifically, the ARPANET 
would take a message and break it up into multiple packets. They will not be delivered until all of its packets 
were collected at the destination. So when packets arrive at their destination, you have to put aside some buffer 
space to reassemble them. BBN made the calculation and said, q._ook, if I've got this many buffers, what is the 
19 
------------------------------<page break>-----------------------------
maximum number of messages I can reassemble at the same time?' Well, the smallest multi-packet message 
is two packets. One packet doesn't have to be reassembled. So they took the size of the storage and divided 
by two to determine how many messages could be in the process of reassembly. And that is how many pointers 
they installed in their software. A few years later, in 1973, they increased the storage by a factor of two, and they 
forgot to increase the number of pointers. Who would ever think of that. So they could allocate space, but could 
never point to it again! CRASH! It turns out that the measurement messages are two packets long. So a pile 
of them came into our switch, took us down as well as the rest of the network. A similar thing happened just 
last fall to AT&T. So the UCLA-BBN relationship was one of guarded respect. 
O'NEILL: When you had problems, or when you could see that there was a problem with the network, or when 
you were trying to get something accomplished, did you go through the ARPA office - through Larry Roberts? 
Or did you deal with BBN directly? 
KLEINROCK: We dealt with BBN directly. When we had a problem with BBN, we complained to Larry and 
he would step in and make sure things were fixed up. It was not a formal relationship that required all kinds 
of paperwork to go back and forth. It was peers, and researchers, and developers. It was a friendly and efficient 
environment in that sense. 
O'NEILL: I'm tryiag to get a better feel for how Roberts influenced the development of the ARPANET. 
KLEINROCK: He was dominant. In the sense that he set goals, set things in motion, and protected us from 
all the vagaries of the ARPA structure on the other side of him. He managed to keep my side simple - he was 
a master at that. Have you met Larry? 
O'NEILL: No. 
------------------------------<page break>-----------------------------
KLEINROCK: Larry is about my age, in his early f'ffties. He's one of the smartest guys I know. But he 
wouldn't always f'fil in all the blanks when he was saying something. So people sometimes had a hard time 
following him. The things he accomplished while at ARPA were phenomenal! 
TAPE 2/SIDE ONE 
KLEINROCK: The ARPANET development took place during Larry's tenure as did part of the packet satellite 
experiment. Somewhere around that time Bob Kahn came in. It may have been at the end of the packet 
satellite effort. Certainly the packet radio project was pretty much Bob's baby as were many things beyond that. 
You mentioned SRI. SRI was an important member of the community;, as you see, they were the second node 
in the ARPA network. They served as Network Information Center, NIC. We did not have much of a 
relationship with them. Documentation went through them, as did the network RFC notes. BBN was the 
principal player so far as we were concerned in the ARPANET. There's a bit of an anecdote here. It has to 
do with the arrival of the IMP at UCLA. It was supposed to be a week late because they were slow at BBN - 
we heaved a sigh of relief because we needed the time. They put the damn thing on an airplane and air 
freighted it out here - it appeared on time! That next morning everybody was present here. My staff, the 
computer science chairman, the school of engineering administration, somebody from the chancellor's office, 
somebody from AT&T long lin es, local telephone company people, Honeywell, the DARPA guys, the BBN guys - 
and everybody was ready to point the accusing finger! Fortunately that first day we got bits moving back and 
forth properly. The next day messages were moving back and forth. I have to give credit to BBN; they did a 
superb job. Our end worked well; their end worked well. It went off with no hitches. 
O'NEILL: So on the same day that it was delivered... 
KLEINROCK: We were moving bits back and forth. We had a guy here, Mike Wingfield, who was in charge 
of our side of the hardware interface; he was excellent. We have a huge number of photographs with the key 
21 
------------------------------<page break>-----------------------------
people in them, if you are interested in that; they came back for our ACT ONE symposium. There were a lot 
of people involved in this project. 
O'NEILL: What kind of relationship did you have, if any, with the wider host community? 
KLEINROCK: We at UCLA were charged with making the host-to-host software. Steve Crocker was in charge 
of that effort. It was supposed to be completed by the summer of 1970; it started in the summer of 1969. 
O'NEILL: This is the Network Working Group? 
KI.FINROCK: That's right. A bunch of graduate students were given this job. It was not until October 1971 
when the specification of the host-to-host protocol was complete and it began to be installed. It was more than 
a year late. It was a much harder project than we thought, and it was done not only locally at UCLA, but among 
various hosts that were involved, including people at the University of Utah and SRI. So the way the ARPANET 
was used until October 1971 was not easily, but it was used mainly as people migrated from one site to another 
(typically when they changed jobs). They wanted to use the software back at their old hosts, and they knew how 
to use it; that generated a lot of early use. The host-to-host interface was awful to begin with. Before the 
network was installed, I interacted with the early host sites in the following way (as reported in one of my early 
published papers): We tried to get a feeling for how much traffic would be placed on the network between 
different nodes. So I called the individual sites and I said, for example, 'MIT, how much traffic are you going 
to send to Illinois?' And the principal investigator said, 'I don't know;, I have no idea how much I'm going to 
send.' v'ill you send, like, three teletypes' worth?' and they said, 'Yes, three teletypes' worth.' So I collected 
all this data, and I published a paper with that particular traffic matrix. It had no relationship to reality whatever. 
There was an interesting issue here that affected the architecture of the ARPANET. When Larry Roberts and 
Bob Taylor announced this network, most of the hosts did not want to participate. They considered this an 
impediment. Attachment to the network implied that they would have to implement some host software; that 
------------------------------<page break>-----------------------------
is, they would have to do major surgery on their operating system to get this network to talk to them. Once the 
host-to-host protocol was specified that implied they had to implement it on their machine. Now d they had the 
same kind of machine, like a PDP-10, that someone else had, then fine, but d they had somethin E else like an 
ILLIAC IV, they had to do it themselves. Even early on, people were not happy about this. They didn't want 
people using their machine, etc. So Larry had to do a sales job and convince people that there was good reason 
to implement it. As a result, some of the ways in which the network was specified was done such that it would 
not impact the user very much. That made it more complicated for the internal operations of the network to 
do things in the network that should have been done outside the network. In general, people said, 'No, I don't 
want to participate.' And we managed to get most of them to participate eventually, as you can see. 
O'NEILL: 
feeling? 
KLEINROCK: 
Did they come right out and say, 'No, we don't want to do this,' or was it just kind of a general 
Well, Larry couldn't say, 'Thou shalt.' 
O'NEH.I.: But he was controlling the money. 
KLEINROCK: Yes, but he wasn't that crude. He encouraged them. Larry was not dictatorial that way;, he 
was decisive, but he wouldn't really force people. He would cajole them, strongly. They soon began to realize 
that there was a benefit. You see the biggest surprise about the ARPA network use was e-mail. It was an ad 
hoc add-on by BBN, and it just blossomed. And that sucked a lot of people in. It still is the biggest use of 
networks today. Orionally, the network was supposed to provide resource sharing, not even data shax'ing. For 
example, you would log on to Utah to use their graphics capability there. At one time it was thought maybe you 
could import the software to your machine and run it locally. But the original concept was that you would do 
resource through the network - that's not really what happened. What happened was we used it as a 
communications medium for access to data, as opposed to access to programs. Early on, Rand was using the 
------------------------------<page break>-----------------------------
machine here at UCLA. The guys at Rand had earlier been carrying decks of cards over to UCLA to run their 
weather simulation and climate simulation on our big 360/91 at the time. Once the network came on line, they 
generated a lot of traffic going back and forth; it was really good stuff, too. So the community of people using 
the network was the scientists as well as the computer-types who were using computing for computing's sake, 
and not for computer science's sake. 
O'NEILL: Were you using the network early on as well? Were you using the e-mail and seeing these benefits 
for yourself?. 
KLEINROCK: E-mail, yes. I wasn't using the external graphics. Neither did I use resource sharing. I used 
it as a vehicle for studying networks. 
O'NEILL: I want to understand what was meant by resource sharing. Was it the same as distributed s 
processing? Was that a goal? 
KLEINROCK: That is a concept nowadays. Actually it was and was not. Let me preface that by asking, are 
you familiar with the October 1972 Ik,,public demonstration of the ARPANET? 
O'NEILL: Yes. 
KLEINROCK: It was announced months ahead of time. All of us were told we were going to have this 
demonstration. DARPA installed an IMP in a hotel in Washington, D.C. and ran in some lines. Everybody was 
encouraged to create some demonstration packages, and we did as well. That caused lots of good things to 
happen in the ARPA network. It generated lots of new uses of the ARPA network, just for that demo. One 
of the things that was demonstrated there was a distributed air traffic control system. The idea was there would 
be a bunch of computers in the network that would be simulating air traffic control operation in their physical 
24 
------------------------------<page break>-----------------------------
region. For example, MIT would be doing Boston, and some Washin,on machines would do Washington and 
so on in different regions. The idea was there would be some air traffic on one of the computers. As this plane 
moved out of its region, it would be picked up by another computer's air traffic control, data would be 
exchanged, and collectively they would be managing this whole air space over a large region. This package also 
had the ability to freeze the simulation, take the program on machine A, squeeze it over to machine B, and then 
continue the simulation. There were real/y some sophisticated thing. going on there. I remember one of the 
demos was really interesting. In this demo, you would sit down in Washlnon at a teletype, logon to a machine 
at BBN, pull up some source code, ship it over to a machine at UCLA across the country, compile and execute, 
and bring back the results to be printed on the teletype right next to you in Washington. Now at the same time 
there were some artificial intelligence demonstrations going on. There was a thing called Turtle running around 
the ground, bumping into things and avoiding them - MIT had that. Jori Postel sat down to demonstrate this 
thln logon to BBN, shift it over to UCLA, compile and execute, and then send it back to print. Nothing 
happened! He couldn't figure out what was wrong. He kept looking around. Then he found Turtle on the 
floor, and Turtle was jumping around... It turns out that the output was going to Turtle - they had messed up 
a connection. There's your output jumping around on the floor. (Laugh) But the point is it was a great demo. 
People were pulled out of the hal/way, handed a handbook, and told, 'Sit down, we'll help you use the 
ARPANET,  and they could. So that was another case where a lot of hosts got involved, and distributed 
processing came about. But that was not the main purpose. The main purpose was to prove networking. And 
from Larry's point of view as the person funding all of this, he was pleased to get some user sharing. And Larry 
did a paper early on to show the cost effectiveness of the ARPANT. He said if you had to replicate all of this 
at every site, it would cost you millions and millions; and here was a network providing this capability at far less 
cost; therefore we win. This was a false argument. No one was going to reproduce the ILLIAC IV at all of 
these places. 
------------------------------<page break>-----------------------------
O'NEILL: You mentioned that Roberts shielded you and the other contractors from a lot of what was going 
on behind the scenes at ARPA, or that's how you felt. Did you have any exposure to the military aspect of 
ARPA? Did you ever feel like you had to have concern for DoD interests.* 
KLEINROCK: Yes. Every time I wrote a proposal I had to show the relevance to the military's applications. 
The whole packet satellite, and certainly the packet radio, was dearly military - soldiers running around, tanks, 
troop carriers, aircraft, submarines. Yet, we never did any military work. We realized there was an application 
there, but that didn't drive our research or our experimentation. It may be that people like Rockwell, who 
developed the packet radio, were told, 'It has to be small and lightweight, and low power' and all the rest. In 
fact, the survivable aspects of it, I'm sure, were dictated by the military. You could chop pieces out of it and 
still make it work. But we had only to identify these applications, and to point out how this body of knowledge 
would be used for that. It was not at all imposed on us. But it de'mitely colored our proposals. I mean, we 
were well aware that we had to put that in in order to get funding. 
O'NEILL: Was that from the very benning or did that change? Do you recall much difference? 
KLEINROCK: It got worse and worse as time went on. More and more - instead of worse and worse. 
O'NEILL: You mentioned that some of the early people who were to be hosts were negative about the idea 
of sharing their resources. Were you exposed to other negative reactions to the network? For example, people 
have mentioned the phone company being very negative. 
KLEINROCK: Oh, yes, that was clear. It has been said that the telephone industry, or the communications 
industry, had absolutely nothing to do with the development of the ARPA network, of packet networks. To first 
order that's correct. You know, IBM was one of the people at that 1967 meeting, as I recall. They sent a fellow 
named Doug McElroy who was a good man, interested, with good ideas. And then IBM dropped out, and I 
------------------------------<page break>-----------------------------
think the reason was that they wanted to bid on this thin and they couldn't be part of the inner circle and also 
bid. They never appeared again in the community at all. Of course, they went SNA. Also, AT&T was not 
involved as an organization. I remember in the late 1960s what was going on at the various computer 
conferences. There would always be a panel consisting of both industries, data communications and data 
proceing. There would be a debate between those two. The computer guys would say, 'Communication guys, 
will you please give us good data communications.' The communications guys would turn around and say, %Vhat 
are you talltin E about, the United States is a copper mine. You've got wires all over the place; use them.' The 
computer guys would say, 'No, you don't understand. It takes haft a minute to set up a call, and your charge 
is for a minimum of three minutes. All I want to do is send a hundred milliseconds of data.' These guys would 
turn around back to the computing guys and say, 'Go away little boy, there's no revenue there.' So the little boys 
went away, and they created packet switching. There was no revenue, and in some sense, there still isn't 
sufficient revenue. The way the communication carriers are going to make money out of data is to ditize voice 
and to offer new data service. They've done the former by packefizing voice. Now they have made a distal 
network out of the old analog network, but were reluctant to do so. In fact, just look at the history. They didn't 
get into networking until GTE bought up Telenet in 1979. In April 1978, AT&T was supposed to develop 
somethin E called the Bell Data Network. They never did. In 1979 they started talking about AIS, Advanced 
Information Service, another network they never made. Then finally in 1982, they came out with Net 1000. In 
1986 they closed down Net 1000; they lost a billion dollars on that effort. Now they have Accunet and more. 
It took them decades to come up to the technology that the data processing guys developed in the ARPANET. 
O'NEILL: Was the term 'packet switching' used all along? There were some early papers where they refer to 
message switching although the idea of the packet was there. 
KLEINROCK: The idea of the packet was there from the initial days of the ARPANET. It's also implicit in 
my 1964 book, 'Communication Nets; Stochastic Message Flow and Delay.' 
27 
------------------------------<page break>-----------------------------
O'NEILL: Maybe we should wrap this up because I know you have an appointment coming up. First let me 
ask you just one or two questions about Act One. I did see the people that you had invited. What was the goal 
of Act One? And did you invite the people that you thought had been the main players or are currently the 
active players? How did you decide who to invite? 
KLEINROCK: There were two reasons for ACT ONE. First, it was the twentieth anniversary of the 
ARPANET, and the ARPANET we knew was going to be killed. So we felt we should commemorate that event. 
Second, UCLA did want to start a symposium series, and it was a great way to launch it. That was the 
motivation. Steve Crocker was one of the people who originated the idea of a commemoration. He made the 
point, qook, ARPANET is going away;, it's twenty years; let's do something.' So he was one of the prime 
movers. So he, Jerry Estrin, and I were the key people in bringing about the ACT ONE symposium. Jerry and 
I turned out to be co-chairmen because we are with the university, but Steve was one of the inner members. 
And we invited the people to each session that we thought were appropriate: the ARPANET lxistory with the 
ARPANET history guys, today's technology, and the future. We asked various of the players who they thought 
should be invited. And some people could make it, some. people couldn't. We ended up with excellent 
representation. In fact, in this case, the audience was every bit as qualified as the speakers. We had a lot of 
the guys out there - the old guys and the new guys. I don't know if you heard, but the venue, the whole flavor, 
was first rate... Did you know there was a lot of poetry at ACT ONE? 
O'NEILL: Alex McKenzie had some of it that I took a look at. 
KLEINROCK: It was just a fun thing. It was done with class; it was a class act. I've got the poetry somewhere, 
because I wrote some of it. Vint wrote some, and Barry Boehm wrote some. 
O'NEILL: Do you have any general comments about the ARPANET, or about its development, that you would 
like to add? 
------------------------------<page break>-----------------------------
KLEINROCK: Yes. I think that DARPA IPTO (now ISTO) was a prime mover for the United States in the 
advancement of computer technology through advanced thinking and... what shall I say... heroic funding of the 
things they thought were worthwhile. Their motto was, rtigh risk, high payoff.' That's exactly what they did 
with a number of projects. You named them before: time-sharing, networking, AI, a few other areas. 
Networking was one of their major successes. They backed it fully;, lots of money, lots of freedom in terms of 
what we were doing, really advanced technology. It was a great, great experiment - I can't applaud them more. 
It was one of the great experiments in science, I think. It completely changed the way things are going on now - 
commerce, government, industry, science, etc. Who to thank? The key players at DARPA for sure. But I think 
the orinators of the whole ARPA concept - Licklider being one of them. The concept of qet's fund the kinds 
of projects and the people that we believe know how to carry out these projects well.' And they cut through a 
lot of red tape. They didn't go through all the peer reviews; they just bet on people that they had confidence 
in. I think it worked real well. It was a heroic kind of thing. On a scale of one to ten I would give them a ten. 
Of course, there were some bad things along the way. Whenever you are aggressive and advanced, there is going 
to be something and somebody left out; you can't help that. 
O'NEILL: So that was the price that you paid for the success. 
KLEINROCK: It was not done in a democratic way. Not every step was carefully judged, 'Should we do this, 
that, or the other thing? Let's get thirty people to vote.' Some guy said, q_ook, I believe that is right - let's do 
it.' And they did it. And the money was forthcoming. There were a lot of hurdles, telephone company hurdles, 
government hurdles, university hurdles - all the way. And that was not only for the ARPANET project but also 
other projects as well. 
O'NEILL: You mentioned that Roberts was shielding you, or that you felt he was shielding you, from what was 
going on behind the scenes. Do you know what was going on? What he was doing? Or did you just see the 
end result of that - that you didn't have to hassle with it? 
29 
------------------------------<page break>-----------------------------
KLEINROCK: Now for BBN. I characterize the BBN relationship this way. BBN, as 
were most software groups at that time, were largely uninterested in 
performance. That's an exaggeration. But by and large a programtact, a software guy, 
simply wants to get a piece of software that works. That's hard enough. Whether it works 
efficiently or well is not usually the issue; they won-y about that maybe way down the line. 
So BBN was not focusing on efficiency or performance but on making it work and 
worrying about the logic. So we had a running battle with them in terms of showing them 
how to improve performance. Another student of mine (Gm'y Fultt) and I showed them 
early on that their routing procedure was infehor to two others we developed. We did this 
in 1970. Finally in 1979 they implemented something very much like one of our 
procedures and then some years latcr another piece of it. BBN monitored the 
network. They measured traffic, and they measured line failures, and IMP 
failures. We were doing things like measuring response times (which they 
were not doing) and stressing the network by generating extra network 
traffic to see what would happen. We would, almost on demand, break the 
network. There were faults in the software, logic faults. So you can see 
BBN was not very happy with us showing up their faults and telling them 
to fix them. Sooner or later they did. Any large system is going to contain faults. But 
we became experts in finding these things, and I document a whole bunch of them in here. 
One time we were making measuremcnks, which means we turned on the IMP software 
which made measurements and which were sent back to us at the Network Measurement 
Center. As a result of that the network kept crashing. So BBN called us up and said, 
"You're crashing the network!" And we said, "We need to make measurements." And we 
found out why. Are you a technical person at all? 
O'NEILL: Yes, somewhat. 
KLEINROCK: Well, the IMP had only so much storage, initi',dly a very limited amount of 
storage. And some of it was set aside to reassemble packets into messages - the 
reassembly function. Specifically, the ARPANET would take a message and break it up 
into multiple packets. They will not be delivered until all of its packets were collected at the 
destination. So when packets an'ive at their destination, you have to put aside some buffer 
space to reassemble them. BBN made the calculation and said, "Look, if I've got this 
many buffers, what is the maximum number of messages I can reassemble at the same 
time?" Well, the smallest multi-packet message is two packets. One packet doesn't have to 
be reassembled. So they took the size of the storage and divided by two to determine how 
many messages could be in the process of reassembly. And that is how many pointers they 
installed in their software. A few years later, in 1973, they increased the storage by a 
factor of two, and they forgot to increase the number of pointers. Who would ever think of 
that. So they could allocate space, but could never point to it again! CRASH! It turns out 
that the measurement messages are two packet.s long. So a pile of them came into our 
switch, took us down as well as the rest of the network. A similar thing happened just last 
fall to AT&T. So the UCLA-BBN relationship was one of guarded respect. 
------------------------------<page break>-----------------------------
Katie's Well 
X-Sender: walden@labs-n.bbn.com 
Date: Fri, 83 Feb 1995 09:12:00 -0600 
To: kahieh@bga.com 
Subjecn: ccTents on ien's co?rents 
<<I put my comments on Len's comments in double angle brackets. 
You should also check whan i said about bugs in the IMP in 
my interview by Judy C'Neii.>> 
Now for BBN. I characterize he BBN relationship this way. BBN, as 
were most software groups at that time, were largely uninterested in 
perfomance. 
<<BBN was no just a software group. BBN was strong in hardward, 
software, theory and systems design and implementation. For 
instance, Ornstein, Barker and Heart were hardware people. Crowther 
Cosell and I were software people. Kahn and Crowther were powerful 
with theory. All of us had successfully built complex systems. 
Our greauesu szrength then and for years to come was the ability to 
There are 141 lines left (16%). Press <space> for more, or 'i' to return. 
Message 1/93 From David C. Walden Page 2 
Our greatest strength then and for years to come was the ability to 
find the bose mix between hardware and solEware.>> 
That's an exaggeration. But by and large a programmer, 
a software guy, simply wants to get a piece of software that works. 
That's hard enough. Whether it works efficiently or well is not 
usually the issue; they worry abouE uhat meybe way down the line. 
So BBN was not focusing on efficiency or performance bun on making 
work and worrying about the logic. 
<<It appears to me tha2 one needs to define what one means 
by "perfomr. arce." ,,en Len says "performnce," he appears to 
be talking about the efficiency and correctness network wide 
algorinhras. i'll come back to these in a minute, but we were 
also very concerned with "performance" in the sense of fitting 
a lot of complexity into a little box and making the little box 
rmn very very fast. We were also concerned with efficiency of 
nonwork wide aigoriEhras, and this actually was an odds with us 
wanting te make some of the changes that would lead to more 
correct network wide algorithms. Many software groups have 
the characteristics Len describes, but our 3 person group 
There are 121 lines left (28%). Press <space> for more, or 'i' o return. 
Message 1/93 From David C. Walden 
Page 3 
the characteristics Len describes, but our 3 person group 
was enormously focused on program (it not always network) 
performance, and in fact we had calculated out the program 
perfo.mnce in advam. ce of wriEing i. 
Also, BBN didP.'t speak with one voice about the issues 
Len brings up. For instance, you should get Kahn's views 
on the situation. He views himself as having been in a 
running battle with Crowther about the correctness of the 
network algorithms. Also, farily early, before he work 
by Len's studonus, I hink, K and I rn the famous 
tests at UCLA that showed some of the problems with the 
initial network algorithms and which Kahn views' as having 
proved that Crowther should have listened to him better. 
------------------------------<page break>-----------------------------
Katie's Well 2/3/95 
After that Kahn and Crowther wrote a report describing 
the problems.>> 
So we had a running battle with 
them in terms of showing them how to improve performance. Another 
student of mine (Gary Fui%t) nd i showed thr. early on tRat their 
routing procedure was inferior o two others we developed. We did 
There are !01 lines left (40%). Press <space> for more, or 'i' to return. 
Message 1/93 From David C. Walden 
Page 4 
routing procedure was inferior to two others we developed. We did 
this in 1970. Finally in 1979 they implemented something very much 
like one of our procedures and then some years later another piece of 
it. BBN monitored the network. They measured traffic and they 
measured line failures, and iMP failures. We were doing things like 
measuring response times (which nkey were not doing) nd stressing the 
network by generating extra network traffic to see what would happen. 
We would, almost on demand, break the network. There were faults in 
the software, logic faults. So you can see BBN was not very happy 
with us showing up their faults and telling them to fix them. Sooner 
or later they did. Any large system is going to contain faults. But 
we became experts in finding these things, and I document a whole 
bunch of them in here. One time we were making measurements, which 
means we turned on the IMP software which made measur.ents and which 
were senn back no us at the Network Measurer. ent Center. As a result 
of that the network kept crashing. So BN called us up nd said, 
"You're crashing the network!" And we said, "We need to make 
measurements." And we found out why. Are you a technical person at 
all? 
<<I don't view there as having been much of a running battle, 
There are 81 lines left (52%). Press <space> for more, or 'i' to return. 
Message 1/93 From David C. Walden 
Page 5 
<<I don't view there as having been mcch of a r'ning battle, 
and i was responsible for operation of the network for mach 
of this time. In fact, from the earliest days we were also 
writing papers and reports on problems we discovered or fixes 
to problems that others discovered. There could have been 
some tension between our job of keeping the network up and running 
and them scheduling time to test it and possibly break it, 
but i always felt that that was mostly worked out easily enough. 
As Len describes, they had the job from ARPA of doing network 
measurement, nd we mostly spent our time doing the job that we 
had from ARPA, which was a different job. We got their reports, 
tried to figure out ways to correct the problems they discovered, 
we used some of their suggested methods and didn't use some 
others. My view is that it was all pretty collaborative and 
together we did some pretty amazing stuff. 
I do believe to this day that some of the work by Let and 
his people was overly theoretical, i.e., i9nored too mny 
real life issues, and that .7y experience with Let 
was one of the things that made me really happy to 
There are 61 lines left (63%). Press <space> for more, or 'i' to return. 
Message 1/93 From David C. Walden 
Page 6 
was one of the things that made me really happy to 
be in indust!z where people care less about who gets credit 
for what is done and more about getting something done. Seeing 
------------------------------<page break>-----------------------------
Katie's Well 
the academic inclination hold back telling us about things until 
they were published, when we would tell them anything as soon as 
we knew i was a real eye-opener to me. Of course, Len 
m. ay see it exactly in reverse. 
As ! said above, in general, i have the utmost respect for 
he work Len and his people did, and hink that our implementation 
inclination combined wih their empirical inclination produced 
as beer resul Ehat either of us would have produced alone. 
Speaking for me personally, I never fel very involved in ny 
of he tension Len describes. Mostly i just hopped along telling 
nyone who would listen wha we were doing, listening to whatever 
anyone else said, and rying to make it all better as best 
we could.>> 
O'NEILL: Yes, somewhat. 
There are 41 lines left (7%). Press <space> for more, or 'i' to return. 
Message 1/93 From David C. Walden 
Page 7 
KLEINROCK: Well, he IMP had only so much sorage, initially a very 
limited amoun of 
sorage. Amd some of it was set aside to reassemble packets into 
messages - the reassembiy f,mncion. SDecificai!y, nhe RPkNT would 
take a message nd break it up into multiple packets. They will not 
be delivered until all of is packets were collected at the 
desEination. So when packets arrive at their destination, you have o 
put aside some buffer space to reassemble Ehem. BBN made the 
calculaEion and said, "Look, if I've got this many buffers, what is 
the maximmm number of messages I can reassemble aE the same time?" 
Well, the smailesn multi-packen message is two packets. One packet 
doesn't have to be reassembled. So they took the size of the storage 
and divided by two to determine how many messages could be in Ehe 
process of reassembly. And hat is how many pointers they installed 
in their software. A few years later, in 1973, they increased Ehe 
storage by a factor of two, and they forgo Eo increase the number of 
pointers. Who would ever think of hat. So they could allocate 
space, but could never poin to it again! CRASH! It urns ou tha 
the measuremen messages are Ewo packets long. So a pile of them came 
inEo our switch, Eook us do as well as the rest of he network. A 
There are 21 lines left (87%). Press <space> for more, or 'i' to return. 
Message 1/93 From David C. Walden 
into our switch, took us down as well as the rest of the network. 
similar uhing happened jusn lasm fall to AT&T. So the UCLA-BBN 
relationship was one of guarded respect. 
A 
<<I don' actually remember the instance Len describes, but such 
bugs do get put into programs. Being really experienced 
implementors, being caught with bugs wasn't very threatening 
o us. It was more of a "damn, made another big mistake 
in public again, big sigh, well, how do we fix his one." 
My respec for UCLA was never guaried. i was also enormously 
impressed wih Kleinrock, Cerf, Crocker, Postel, Wingfield, 
Estrin, Naylor, Fulz, Lam, Gerla, ec. ec. etc. and I loved 
having the opportunity to work wih them.>> 
Page 8 
------------------------------<page break>-----------------------------